Export models from https://github.com/voicekit-team/T-one to sherpa-onnx (#2571)

This PR exports models from the T-one repository (https://github.com/voicekit-team/T-one) to sherpa-onnx format, creating a complete pipeline for Russian speech recognition using streaming CTC models.

- Adds scripts to download, process, and test T-one models in sherpa-onnx format
- Creates GitHub workflow for automated model export and publishing
- Updates kaldi-native-fbank dependency to version 1.22.1

+name: export-t-one-to-onnx
+
+on:
+  workflow_dispatch:
+
+concurrency:
+  group: export-t-one-to-onnx-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  export-t-one-to-onnx:
+    if: github.repository_owner == 'k2-fsa' || github.repository_owner == 'csukuangfj'
+    name: export t-one
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest]
+        python-version: ["3.10"]
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install Python dependencies
+        shell: bash
+        run: |
+          pip install onnx==1.17.0 onnxruntime==1.17.1 soundfile librosa kaldi_native_fbank "numpy<2"
+
+      - name: Run
+        shell: bash
+        run: |
+          cd scripts/t-one
+
+          wget https://raw.githubusercontent.com/voicekit-team/T-one/refs/heads/main/LICENSE
+          ./run.sh
+
+          d=sherpa-onnx-streaming-t-one-russian-2025-09-08
+          mkdir $d
+          cp -v ./tokens.txt $d
+          cp -v ./model.onnx $d
+          cp -v ./russian_test_short_from_t_one.wav $d/0.wav
+          cp -v ./LICENSE $d
+          cp -v ./README.md $d
+
+          ls -lh $d
+
+          tar cjfv $d.tar.bz2 $d
+
+          ls -lh $d.tar.bz2
+
+          mv $d.tar.bz2 ../..
+          mv $d ../..
+
+      - name: Publish to huggingface
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        uses: nick-fields/retry@v3
+        with:
+          max_attempts: 20
+          timeout_seconds: 200
+          shell: bash
+          command: |
+            git config --global user.email "csukuangfj@gmail.com"
+            git config --global user.name "Fangjun Kuang"
+
+            rm -rf huggingface
+            export GIT_LFS_SKIP_SMUDGE=1
+            export GIT_CLONE_PROTECTION_ACTIVE=false
+
+            m=sherpa-onnx-streaming-t-one-russian-2025-09-08
+
+            git clone https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/$m huggingface
+            cd huggingface
+            git fetch
+            git pull
+            echo "pwd: $PWD"
+            ls -lh ../$m
+            git lfs track "*.wav"
+
+            rm -rf ./*
+
+            cp -v ../$m/* ./
+
+            git lfs track "*.onnx"
+            git add .
+
+            ls -lh
+
+            git status
+
+            git commit -m "add models"
+            git push https://csukuangfj:$HF_TOKEN@huggingface.co/csukuangfj/$m main || true
+
+            cd ..
+
+      - name: Release
+        uses: svenstaro/upload-release-action@v2
+        with:
+          file_glob: true
+          file: ./*.tar.bz2
+          overwrite: true
+          repo_name: k2-fsa/sherpa-onnx
+          repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }}
+          tag: asr-models

@@ -148,3 +148,4 @@ voices.bin
 kitten-nano-en-v0_1-fp16
 kitten-nano-en-v0_1-fp16
 *.egg-info
 *.egg-info
 *.jar
 *.jar
+vocab.json

 function(download_kaldi_native_fbank)
 function(download_kaldi_native_fbank)
   include(FetchContent)
   include(FetchContent)
 
 
-  set(kaldi_native_fbank_URL   "https://github.com/csukuangfj/kaldi-native-fbank/archive/refs/tags/v1.21.3.tar.gz")
-  set(kaldi_native_fbank_URL2  "https://hf-mirror.com/csukuangfj/sherpa-onnx-cmake-deps/resolve/main/kaldi-native-fbank-1.21.3.tar.gz")
-  set(kaldi_native_fbank_HASH "SHA256=d409eddae5a46dc796f0841880f489ff0728b96ae26218702cd438c28667c70e")
+  set(kaldi_native_fbank_URL   "https://github.com/csukuangfj/kaldi-native-fbank/archive/refs/tags/v1.22.1.tar.gz")
+  set(kaldi_native_fbank_URL2  "https://hf-mirror.com/csukuangfj/sherpa-onnx-cmake-deps/resolve/main/kaldi-native-fbank-1.22.1.tar.gz")
+  set(kaldi_native_fbank_HASH "SHA256=b292ddd1fa121f28371d11c14dd016c59c54b3f0dbb2bb2cfdc82d562564d0f5")
 
 
   set(KALDI_NATIVE_FBANK_BUILD_TESTS OFF CACHE BOOL "" FORCE)
   set(KALDI_NATIVE_FBANK_BUILD_TESTS OFF CACHE BOOL "" FORCE)
   set(KALDI_NATIVE_FBANK_BUILD_PYTHON OFF CACHE BOOL "" FORCE)
   set(KALDI_NATIVE_FBANK_BUILD_PYTHON OFF CACHE BOOL "" FORCE)
@@ -12,11 +12,11 @@ function(download_kaldi_native_fbank)
   # If you don't have access to the Internet,
   # If you don't have access to the Internet,
   # please pre-download kaldi-native-fbank
   # please pre-download kaldi-native-fbank
   set(possible_file_locations
   set(possible_file_locations
-    $ENV{HOME}/Downloads/kaldi-native-fbank-1.21.3.tar.gz
-    ${CMAKE_SOURCE_DIR}/kaldi-native-fbank-1.21.3.tar.gz
-    ${CMAKE_BINARY_DIR}/kaldi-native-fbank-1.21.3.tar.gz
-    /tmp/kaldi-native-fbank-1.21.3.tar.gz
-    /star-fj/fangjun/download/github/kaldi-native-fbank-1.21.3.tar.gz
+    $ENV{HOME}/Downloads/kaldi-native-fbank-1.22.1.tar.gz
+    ${CMAKE_SOURCE_DIR}/kaldi-native-fbank-1.22.1.tar.gz
+    ${CMAKE_BINARY_DIR}/kaldi-native-fbank-1.22.1.tar.gz
+    /tmp/kaldi-native-fbank-1.22.1.tar.gz
+    /star-fj/fangjun/download/github/kaldi-native-fbank-1.22.1.tar.gz
   )
   )
 
 
   foreach(f IN LISTS possible_file_locations)
   foreach(f IN LISTS possible_file_locations)

+# Introduction
+
+This folder contains scripts for exporting models from
+https://github.com/voicekit-team/T-one
+to sherpa-onnx.

+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+
+import onnx
+
+
+def main():
+    meta_data = {
+        "model_type": "t-one",
+        "language": "Russian",
+        "version": 1,
+        "maintainer": "k2-fsa",
+        "sample_rate": 8000,
+        "frame_length_ms": 300,  # chunk_duration_ms
+        "state_dim": 219729,
+        "comment": "This is a streaming CTC model for Russian with expected audio sample rate 8000",
+        "url": "https://github.com/voicekit-team/T-one",
+        "see_also": "https://huggingface.co/t-tech/T-one",
+    }
+    model = onnx.load("./model.onnx")
+
+    while len(model.metadata_props):
+        model.metadata_props.pop()
+
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = str(value)
+    print("--------------------")
+
+    print(model.metadata_props)
+
+    onnx.save(model, "./model.onnx")
+
+
+if __name__ == "__main__":
+    main()

+#!/usr/bin/env python3
+# Copyright    2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+import json
+
+
+def main():
+    with open("vocab.json") as f:
+        token2id = json.load(f)
+
+    with open("tokens.txt", "w", encoding="utf-8") as f:
+        for s, i in token2id.items():
+            if s == "|":
+                s = " "
+            if s == "[PAD]":
+                s = "<blk>"
+
+            f.write(f"{s} {i}\n")
+
+
+if __name__ == "__main__":
+    main()

+#!/usr/bin/env bash
+
+set -ex
+
+if [ ! -f ./model.onnx ]; then
+  curl -SL -O https://hf-mirror.com/t-tech/T-one/resolve/main/model.onnx
+fi
+
+if [ ! -f ./vocab.json ]; then
+  curl -SL -O https://hf-mirror.com/t-tech/T-one/resolve/main/vocab.json
+fi
+
+if [ ! -f ./russian_test_short_from_t_one.wav ]; then
+  curl -SL -O https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/russian_test_short_from_t_one.wav
+fi
+
+python3 ./add_meta_data.py
+
+if [ ! -f ./tokens.txt ]; then
+  python3 ./generate_tokens.py
+fi
+
+./test.py  --model ./model.onnx  --tokens ./tokens.txt --wave ./russian_test_short_from_t_one.wav

+#!/usr/bin/env python3
+# Copyright      2025  Xiaomi Corp.        (authors: Fangjun Kuang)
+
+import argparse
+from typing import Tuple
+
+import kaldi_native_fbank as knf
+import numpy as np
+import onnxruntime as ort
+import soundfile as sf
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        help="Path to model.onnx",
+    )
+
+    parser.add_argument(
+        "--tokens",
+        type=str,
+        required=True,
+        help="Path to tokens.txt",
+    )
+
+    parser.add_argument(
+        "--wave",
+        type=str,
+        required=True,
+        help="The input wave to be recognized",
+    )
+
+    return parser.parse_args()
+
+
+class OnnxModel:
+    def __init__(self, filename):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 1
+        session_opts.intra_op_num_threads = 1
+
+        self.session_opts = session_opts
+
+        self.model = ort.InferenceSession(
+            filename,
+            sess_options=self.session_opts,
+            providers=["CPUExecutionProvider"],
+        )
+
+        meta = self.model.get_modelmeta().custom_metadata_map
+
+        self.frame_length_ms = int(meta["frame_length_ms"])
+        self.sample_rate = int(meta["sample_rate"])
+        self.state_dim = int(meta["state_dim"])
+
+    def get_init_state(self, batch_size=1):
+        return np.zeros((batch_size, self.state_dim), dtype=np.float16)
+
+    def __call__(self, x, state):
+        """
+        Args:
+          x: (batch_size, num_samples, 1), int32
+          state: (batch_size, 219729)
+        Returns:
+          log_probs: (batch_size, num_frames, vocab_size)
+          next_state: (batch_size, 219729)
+        """
+        log_prob, next_state = self.model.run(
+            [
+                self.model.get_outputs()[0].name,
+                self.model.get_outputs()[1].name,
+            ],
+            {
+                self.model.get_inputs()[0].name: x,
+                self.model.get_inputs()[1].name: state,
+            },
+        )
+        return log_prob, next_state
+
+
+def load_audio(filename: str) -> Tuple[np.ndarray, int]:
+    data, sample_rate = sf.read(
+        filename,
+        always_2d=True,
+        dtype="float32",
+    )
+    data = data[:, 0]  # use only the first channel
+    samples = np.ascontiguousarray(data)
+    return samples, sample_rate
+
+
+def load_tokens(filename):
+    ans = dict()
+    with open(filename, encoding="utf-8") as f:
+        for line in f:
+            fields = line.strip().split()
+            if len(fields) == 1:
+                ans[int(fields[0])] = " "
+            else:
+                ans[int(fields[1])] = fields[0]
+    return ans
+
+
+def compute_feat(
+    samples,
+    sample_rate,
+    frame_length_ms: int,
+):
+    opts = knf.RawAudioSamplesOptions()
+    opts.frame_opts.samp_freq = sample_rate
+    opts.frame_opts.frame_length_ms = frame_length_ms
+    opts.frame_opts.frame_shift_ms = frame_length_ms
+
+    raw_audio_samples = knf.OnlineRawAudioSamples(opts)
+
+    raw_audio_samples.accept_waveform(sample_rate, samples)
+    raw_audio_samples.input_finished()
+
+    features = []
+
+    for i in range(raw_audio_samples.num_frames_ready):
+        f = raw_audio_samples.get_frame(i)
+        features.append(f)
+
+    return (np.array(features, dtype=np.float32) * 32768).astype(np.int32)
+
+
+def main():
+    args = get_args()
+    print(vars(args))
+
+    model = OnnxModel(filename=args.model)
+
+    samples, sample_rate = load_audio(args.wave)
+    if sample_rate != model.sample_rate:
+        import librosa
+
+        samples = librosa.resample(
+            samples, orig_sr=sample_rate, target_sr=model.sample_rate
+        )
+        sample_rate = model.sample_rate
+
+    # Pad 0.5 seconds
+    samples = np.pad(samples, (0, 4000))
+
+    features = compute_feat(
+        samples=samples,
+        sample_rate=sample_rate,
+        frame_length_ms=model.frame_length_ms,
+    )
+    print(features.shape)
+
+    id2token = load_tokens(args.tokens)
+
+    blank = -2
+    for idx, token in id2token.items():
+        if token == "<blk>":
+            blank = idx
+
+    state = model.get_init_state()
+    token_id_list = []
+    for f in features:
+        log_probs, state = model(f[None, :, None], state)
+
+        max_token_ids = log_probs[0].argmax(axis=-1).tolist()
+        token_id_list += max_token_ids
+
+    unique_ids = []
+    prev = -1
+    for t in token_id_list:
+        if t == blank:
+            prev = t
+            continue
+
+        if t == prev:
+            continue
+
+        prev = t
+        unique_ids.append(prev)
+    text = "".join([id2token[i] for i in unique_ids])
+    print(text)
+
+
+if __name__ == "__main__":
+    main()